xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/3x8c2-minmax-rndnu-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-dup.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld2r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     const int8_t* restrict a,
22     size_t a_stride,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
28 {
29   assert(mr != 0);
30   assert(mr <= 3);
31   assert(nc != 0);
32   assert(kc != 0);
33   assert(kc % sizeof(int8_t) == 0);
34   assert(a != NULL);
35   assert(w != NULL);
36   assert(c != NULL);
37 
38   kc = round_up_po2(kc, 2 * sizeof(int8_t));
39   const int8_t* a0 = a;
40   int8_t* c0 = c;
41   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
42   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
43   if XNN_UNPREDICTABLE(mr < 2) {
44     a1 = a0;
45     c1 = c0;
46   }
47   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
48   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
49   if XNN_UNPREDICTABLE(mr <= 2) {
50     a2 = a1;
51     c2 = c1;
52   }
53 
54   do {
55     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
56     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
57     int32x4_t vacc1x0123 = vacc0x0123;
58     int32x4_t vacc1x4567 = vacc0x4567;
59     int32x4_t vacc2x0123 = vacc0x0123;
60     int32x4_t vacc2x4567 = vacc0x4567;
61 
62     size_t k = kc;
63 
64     while (k >= 16 * sizeof(int8_t)) {
65       const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
66       const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
67       const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
68       const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
69       const int16x4x2_t va10x0 = vld2_dup_s16((const void*)a1);
70       const int16x4x2_t va11x0 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
71       const int16x4x2_t va10x1 = vld2_dup_s16((const void*)a1);
72       const int16x4x2_t va11x1 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
73       const int16x4x2_t va20x0 = vld2_dup_s16((const void*)a2);
74       const int16x4x2_t va21x0 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
75       const int16x4x2_t va20x1 = vld2_dup_s16((const void*)a2);
76       const int16x4x2_t va21x1 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
77       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
78       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
80       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
81       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
82       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
84       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85 
86       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
87       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
88       const int8x8_t va1c0x0 = vreinterpret_s8_s16(va10x0.val[0]);
89       const int8x8_t va1c0x1 = vreinterpret_s8_s16(va10x1.val[0]);
90       const int8x8_t va2c0x0 = vreinterpret_s8_s16(va20x0.val[0]);
91       const int8x8_t va2c0x1 = vreinterpret_s8_s16(va20x1.val[0]);
92 
93       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
94       int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0x0, va1c0x0);
95       int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0x0, va2c0x0);
96       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
97       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
98       vprod1x0123c0 = vmlal_s8(vprod1x0123c0, vb0123c0x1, va1c0x1);
99       vprod2x0123c0 = vmlal_s8(vprod2x0123c0, vb0123c0x1, va2c0x1);
100       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
101       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
102       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
103       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
104       int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0x0, va1c0x0);
105       int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0x0, va2c0x0);
106       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
107       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
108       vprod1x4567c0 = vmlal_s8(vprod1x4567c0, vb4567c0x1, va1c0x1);
109       vprod2x4567c0 = vmlal_s8(vprod2x4567c0, vb4567c0x1, va2c0x1);
110       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
111       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
112       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
113       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
114       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
115       const int8x8_t va1c1x0 = vreinterpret_s8_s16(va10x0.val[1]);
116       const int8x8_t va1c1x1 = vreinterpret_s8_s16(va10x1.val[1]);
117       const int8x8_t va2c1x0 = vreinterpret_s8_s16(va20x0.val[1]);
118       const int8x8_t va2c1x1 = vreinterpret_s8_s16(va20x1.val[1]);
119 
120       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
121       int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1x0, va1c1x0);
122       int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1x0, va2c1x0);
123       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
124       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
125       vprod1x0123c1 = vmlal_s8(vprod1x0123c1, vb0123c1x1, va1c1x1);
126       vprod2x0123c1 = vmlal_s8(vprod2x0123c1, vb0123c1x1, va2c1x1);
127       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
128       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
129       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
130       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
131       int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1x0, va1c1x0);
132       int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1x0, va2c1x0);
133       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
135       vprod1x4567c1 = vmlal_s8(vprod1x4567c1, vb4567c1x1, va1c1x1);
136       vprod2x4567c1 = vmlal_s8(vprod2x4567c1, vb4567c1x1, va2c1x1);
137       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
138       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
139       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
140       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
141       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
142       const int8x8_t va1c2x0 = vreinterpret_s8_s16(va11x0.val[0]);
143       const int8x8_t va1c2x1 = vreinterpret_s8_s16(va11x1.val[0]);
144       const int8x8_t va2c2x0 = vreinterpret_s8_s16(va21x0.val[0]);
145       const int8x8_t va2c2x1 = vreinterpret_s8_s16(va21x1.val[0]);
146 
147       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
148       int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2x0, va1c2x0);
149       int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2x0, va2c2x0);
150       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
151       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
152       vprod1x0123c2 = vmlal_s8(vprod1x0123c2, vb0123c2x1, va1c2x1);
153       vprod2x0123c2 = vmlal_s8(vprod2x0123c2, vb0123c2x1, va2c2x1);
154       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
155       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
156       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
157       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
158       int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2x0, va1c2x0);
159       int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2x0, va2c2x0);
160       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
162       vprod1x4567c2 = vmlal_s8(vprod1x4567c2, vb4567c2x1, va1c2x1);
163       vprod2x4567c2 = vmlal_s8(vprod2x4567c2, vb4567c2x1, va2c2x1);
164       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
165       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
166       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
167       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
168       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
169       const int8x8_t va1c3x0 = vreinterpret_s8_s16(va11x0.val[1]);
170       const int8x8_t va1c3x1 = vreinterpret_s8_s16(va11x1.val[1]);
171       const int8x8_t va2c3x0 = vreinterpret_s8_s16(va21x0.val[1]);
172       const int8x8_t va2c3x1 = vreinterpret_s8_s16(va21x1.val[1]);
173 
174       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
175       int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3x0, va1c3x0);
176       int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3x0, va2c3x0);
177       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
179       vprod1x0123c3 = vmlal_s8(vprod1x0123c3, vb0123c3x1, va1c3x1);
180       vprod2x0123c3 = vmlal_s8(vprod2x0123c3, vb0123c3x1, va2c3x1);
181       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
182       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
183       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
184       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
185       int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3x0, va1c3x0);
186       int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3x0, va2c3x0);
187       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
189       vprod1x4567c3 = vmlal_s8(vprod1x4567c3, vb4567c3x1, va1c3x1);
190       vprod2x4567c3 = vmlal_s8(vprod2x4567c3, vb4567c3x1, va2c3x1);
191       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
192       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
193       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
194 
195       k -= 16 * sizeof(int8_t);
196     }
197 
198     if (k >= 8 * sizeof(int8_t)) {
199       const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
200       const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
201       const int16x4x2_t va10 = vld2_dup_s16((const void*)a1);
202       const int16x4x2_t va11 = vld2_dup_s16((const void*)(a1 + 4)); a1 += 8;
203       const int16x4x2_t va20 = vld2_dup_s16((const void*)a2);
204       const int16x4x2_t va21 = vld2_dup_s16((const void*)(a2 + 4)); a2 += 8;
205 
206       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
207       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
208       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
209       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
210       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
211       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
212       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
213       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
214 
215       const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
216       const int8x8_t va1c0 = vreinterpret_s8_s16(va10.val[0]);
217       const int8x8_t va2c0 = vreinterpret_s8_s16(va20.val[0]);
218 
219       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
220       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
221       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
222       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
223       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
224       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
225       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
226       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
227       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
228       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
229       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
230       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
231       const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
232       const int8x8_t va1c1 = vreinterpret_s8_s16(va10.val[1]);
233       const int8x8_t va2c1 = vreinterpret_s8_s16(va20.val[1]);
234 
235       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
236       const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
237       const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
238       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
239       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
240       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
241       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
242       const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
243       const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
244       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
245       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
246       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
247       const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
248       const int8x8_t va1c2 = vreinterpret_s8_s16(va11.val[0]);
249       const int8x8_t va2c2 = vreinterpret_s8_s16(va21.val[0]);
250 
251       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
252       const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
253       const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
254       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
255       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
256       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
257       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
258       const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
259       const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
260       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
261       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
262       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
263       const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
264       const int8x8_t va1c3 = vreinterpret_s8_s16(va11.val[1]);
265       const int8x8_t va2c3 = vreinterpret_s8_s16(va21.val[1]);
266 
267       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
268       const int16x8_t vprod1x0123c3 = vmull_s8(vb0123c3, va1c3);
269       const int16x8_t vprod2x0123c3 = vmull_s8(vb0123c3, va2c3);
270       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
271       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c3);
272       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c3);
273       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
274       const int16x8_t vprod1x4567c3 = vmull_s8(vb4567c3, va1c3);
275       const int16x8_t vprod2x4567c3 = vmull_s8(vb4567c3, va2c3);
276       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
277       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c3);
278       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c3);
279 
280       k -= 8 * sizeof(int8_t);
281     }
282 
283     if XNN_UNLIKELY(k != 0) {
284       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
285       const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k);
286       const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k);
287 
288       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
289       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
290 
291       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
292       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
293       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
294       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
295       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
296       const int8x8_t va1c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 0));
297       const int16x8_t vprod1x0123c0 = vmull_s8(vb0123c0, va1c0);
298       vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c0);
299       const int16x8_t vprod1x4567c0 = vmull_s8(vb4567c0, va1c0);
300       vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c0);
301       const int8x8_t va2c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 0));
302       const int16x8_t vprod2x0123c0 = vmull_s8(vb0123c0, va2c0);
303       vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c0);
304       const int16x8_t vprod2x4567c0 = vmull_s8(vb4567c0, va2c0);
305       vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c0);
306 
307       if (k > 2 * sizeof(int8_t)) {
308         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
309         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
310 
311         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
312         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
313         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
314         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
315         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
316         const int8x8_t va1c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 1));
317         const int16x8_t vprod1x0123c1 = vmull_s8(vb0123c1, va1c1);
318         vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c1);
319         const int16x8_t vprod1x4567c1 = vmull_s8(vb4567c1, va1c1);
320         vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c1);
321         const int8x8_t va2c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 1));
322         const int16x8_t vprod2x0123c1 = vmull_s8(vb0123c1, va2c1);
323         vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c1);
324         const int16x8_t vprod2x4567c1 = vmull_s8(vb4567c1, va2c1);
325         vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c1);
326 
327         if (k > 4 * sizeof(int8_t)) {
328           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
329           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
330 
331           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
332           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
333           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
334           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
335           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
336           const int8x8_t va1c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va1), 2));
337           const int16x8_t vprod1x0123c2 = vmull_s8(vb0123c2, va1c2);
338           vacc1x0123 = vpadalq_s16(vacc1x0123, vprod1x0123c2);
339           const int16x8_t vprod1x4567c2 = vmull_s8(vb4567c2, va1c2);
340           vacc1x4567 = vpadalq_s16(vacc1x4567, vprod1x4567c2);
341           const int8x8_t va2c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va2), 2));
342           const int16x8_t vprod2x0123c2 = vmull_s8(vb0123c2, va2c2);
343           vacc2x0123 = vpadalq_s16(vacc2x0123, vprod2x0123c2);
344           const int16x8_t vprod2x4567c2 = vmull_s8(vb4567c2, va2c2);
345           vacc2x4567 = vpadalq_s16(vacc2x4567, vprod2x4567c2);
346         }
347       }
348     }
349 
350     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
351     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
352     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
353 
354     vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
355     vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
356     vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
357     vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
358     vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
359     vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
360 
361     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
362     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
363     vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
364     vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
365     vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
366     vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
367 
368     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
369     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
370     vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
371     vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
372     vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
373     vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
374 
375     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
376 #if XNN_ARCH_ARM64
377     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
378     int16x8_t vacc1x01234567 = vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567);
379     int16x8_t vacc2x01234567 = vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567);
380 
381     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
382     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
383     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
384 
385     int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
386     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
387 #else
388     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
389     int16x8_t vacc1x01234567 = vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567));
390     int16x8_t vacc2x01234567 = vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567));
391 
392     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
393     vacc1x01234567 = vqaddq_s16(vacc1x01234567, voutput_zero_point);
394     vacc2x01234567 = vqaddq_s16(vacc2x01234567, voutput_zero_point);
395 
396     int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
397     int8x8_t vout2x01234567 = vqmovn_s16(vacc2x01234567);
398 #endif
399 
400     const int8x16_t voutput_min = vld1q_dup_s8(&params->rndnu_neon.output_min);
401     vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
402     vout2x01234567 = vmax_s8(vout2x01234567, vget_low_s8(voutput_min));
403 
404     const int8x16_t voutput_max = vld1q_dup_s8(&params->rndnu_neon.output_max);
405     vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
406     vout2x01234567 = vmin_s8(vout2x01234567, vget_low_s8(voutput_max));
407 
408     if (nc >= 8) {
409       vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
410       vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
411       vst1_s8(c2 + 0, vout2x01234567);
412 
413       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
414       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
415       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
416 
417       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
418       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
419       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
420 
421       nc -= 8;
422     } else {
423       // Final case where not all of the 8 columns fit in the destination.
424       if (nc & 4) {
425         vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
426         vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
427         vst1_lane_u32((void*) c2, vreinterpret_u32_s8(vout2x01234567), 0); c2 += 4;
428         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
429         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 4);
430       }
431       if (nc & 2) {
432         vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
433         vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
434         vst1_lane_u16((void*) c2, vreinterpret_u16_s8(vout2x01234567), 0); c2 += 2;
435         vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
436         vout2x01234567 = vext_s8(vout2x01234567, vout2x01234567, 2);
437       }
438       if (nc & 1) {
439         vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
440         vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
441         vst1_lane_s8(c2, vout2x01234567, 0);
442       }
443 
444       nc = 0;
445     }
446   } while (nc != 0);
447 }
448